《BGE-Base-Zh-V1.5 端侧使用教程》不同的文本输入输出的向量相识度很高，没法区分

[深圳市进化图灵智能科技有限公司]

 

代码：https://blog.csdn.net/csdnsqst0050/article/details/157651536

环境：

windows on arm64

python x64:

qai_libs 从 `C:\Qualcomm\AIStack\QAIRT\2.41.0.251128` 复制

(C:\workspace\gitee-Omni-note\model_server\venv) PS C:\models\bge-base-zh-v1.5> pip list |findstr qai
qai_appbuilder            2.38.0
qai-hub                   0.30.0

(C:\workspace\gitee-Omni-note\model_server\venv) PS C:\models\bge-base-zh-v1.5> python .\test.py
Initializing Tokenizer...
Allocating Global Input Buffers...
Initializing QNN Context...
     0.0ms [^main][7352][WARNING]  <W> Initializing HtpProvider
   164.1ms [^main][7352][WARNING] Time: Read model file to memory. 0.40
     0.0ms [^main][7352][WARNING]  <W> m_CFBCallbackInfoObj is not initialized, return emptyList
   393.1ms [^main][7352][WARNING] Time: contextCreateFromBinary. 228.69
   414.4ms [^main][7352][WARNING] Time: UnmapViewOfFile. 21.05
   414.7ms [^main][7352][WARNING] Time: model_initialize bge_context 414.62

[Case 1] Input: 这是一段测试文本
Preprocessing text: '这是一段测试文本'
Global Input IDs (first 5): [ 101 6821 3221  671 3667]
   448.6ms [^main][7352][WARNING] Time: model_inference bge_context 28.35
Embedding (Top 10): [-0.03643476 -0.03728764 -0.02388046 -0.01057563 -0.02889536  0.03331324
  0.00216523  0.00452876  0.02087835  0.00425158]
Shape: (768,)
Saved to embedding_a.npy

[Case 2] Input: 今天天气不错
Preprocessing text: '今天天气不错'
Global Input IDs (first 5): [ 101  791 1921 1921 3698]
   482.5ms [^main][7352][WARNING] Time: model_inference bge_context 26.86
Embedding (Top 10): [-0.03643548 -0.03728837 -0.02388093 -0.01056731 -0.02891298  0.0333139
  0.00217274  0.00452885  0.02089581  0.00426872]
Shape: (768,)
Saved to embedding_b.npy

Releasing resources...
     0.0ms [^main][7352][WARNING]  <W> m_CFBCallbackInfoObj is not initialized, return emptyList
   551.7ms [^main][7352][WARNING] Time: model_destroy bge_context 64.33
QNN Context Released.

 

[深圳市进化图灵智能科技有限公司]

@创达-技术-李晋 我使用qnn版本的 qwen3-reranker-0.6B 的，没法排序：https://bbs.csdn.net/topics/620344238

[创达-技术-李晋]

可以先使用https://www.aidevhome.com/?id=58
Qwen3-embedding-0.6B 使用指南
过渡一下，bge分析看是硬件不支持算子，这边解决会给到

[创达-技术-李晋]

@深圳市进化图灵智能科技有限公司 权限问题，重新下载使用

[深圳市进化图灵智能科技有限公司]

@创达-技术-李晋 (C:\workspace\gitee-Omni-note\model_server\venv) PS C:\models\bge-base-zh-v1.5-qnn-8380-V1.1> python .\inference.py 正在对 3 个文本编码（每个启动独立子进程）... Traceback (most recent call last): File "C:\models\bge-base-zh-v1.5-qnn-8380-V1.1\inference.py", line 190, in <module> vecs = embed_batch(texts) ^^^^^^^^^^^^^^^^^^ File "C:\models\bge-base-zh-v1.5-qnn-8380-V1.1\inference.py", line 161, in embed_batch raise RuntimeError( RuntimeError: Worker failed for '这是一段测试文本' after 3 retries (last rc=3221225477):

[创达-技术-李晋]

可以看下
https://www.aidevhome.com/?id=54
重新下载使用BGE模型